Skip to content

other

Logit lens

An interpretability readout that applies a transformer's final layer norm and unembedding directly to a residual-stream activation to read it as a distribution over tokens.

Current clusters